Transcript PowerPoint 2007

Introduction to Big Data
What is “Big Data”?
Reference: http://en.wikipedia.org/wiki/Big_data)
2
The importance of Big Data in
the Data Science equation
• Large data sets are not new (e.g. Energy, Telecomm, etc.)
• When the data itself becomes part of the problem (e.g. pushing existing limits)
• “ Big Data” embodies a set of tools and technologies for dealing with vast
data sets (e.g. capturing, storing, accessing, processing, etc.)
• Increased data volume dictates increased sophistication in the analysis
and use of that data – the foundation of data science.
© 2013- 2
3
Part I:
Characterizing Big Data
4
Data Size
1,024 Zettabytes (2 80 )
≈10 24
Yottabyte
Yottabyte
Zettabyte
Zettabyte
1,024 Petabyte s (260)
≈10 18
Exabyte
Petabyte
Petabyte
1,024 Gigabyte s (240)
≈10 12
1,0 24 Terabytes (2 50 )
≈10 15
Terabyte
Gigabyte
1,024 Kilobytes (220)
≈10 6
1,0 24 Exabytes (2 70 )
≈10 21
Megabyte
Megabyte
Kilobyte
Kilobyte
30
1,0 24 Megabytes (2 )
≈10 9
10
1, 24 or (2 bytes)
3
0 ≈10
5
Data Format/Composition/Mode of Access
Collection of files; local
vs. network/ distributed/ cloud
Collection of records; text / binary;
structured/ semi- structured/
unstructured (data in motion)
(e.g. audio/ video surveillance,
network monitoring, stocks, etc.)
Collection of data types for
representing compound
entities; fixed length vs.
variable length
Examples:
fixed: (name, DOB)
variable: (name, EmpID, WorkHistory)
00000000 to 11111111
File System
Stream
File
Collection of records; text/ binary;
structured/semi- structured/ unstructured
(data at rest)
(e.g. database, image, video, podcast,
CSV, PDF, HTML, books, journals, etc.)
Record
Data
Type
Data Type
Byte (8 Bits)
Bits)
Binary Digit (Bit)
Collection of bytes for representing
simple and complex entities
(e.g. 123, 3.14, ‘A’, “ Hello There!” ,
[27,59,- 18], (“ what” ,” is” ,” big” ,“ data” ))
0 or 1
6
Data’s V- Dimensions
Volume
Variety
Data Size & Growth Rate
Data types
Velocity
Validity
Veracity
Can the elicited results be believed?
Legitimacy of the data sets (governance provisions)?
Speed requirements
Value
What business advantages
can be gleaned?
Cisco Confidential
7
Part II:
Older Methods of Storing Big Data
File Systems
Hierarchical
Database Model
Network
Database Model
Relational
Database Model
Object Database
Model
Object-Relational
Database Model
XML Database
Model
Content
Management
Systems
Data Warehouse
Distributed
Databases
8
File Systems
File Systems
▪
▪
▪
A collection of information that is arranged in a
hierarchy.
▪
A file corresponds to a container for information.
▪
A directory corresponds to a container for files and
directories.
▪
A sub-directory corresponds to a directory that is
nested within another directory.
Operations
▪
Create, Read, Update, Delete, Find, Navigate
▪
operating system commands
▪
Applications
Examples
▪
Computer Operating Systems (DOS, Windows, Mac
OS, Unix, VMS, etc.)
▪
Network File Systems (NFS), Network Attached
Storage (NAS), File Servers, etc.
9
Hierarchical
Database Model
Hierarchical Database Model
▪
▪
▪
A hierarchical database consists of a collection of
records which are connected to one another through
links.
▪
A record corresponds to a collection of fields; each field
contains a single data value.
▪
A link corresponds to an association between exactly two
records.
▪
Examples
▪
IBM’s Information Management System (IMS)
▪
Microsoft Windows Registry
Dummy Node for Records of type A
Links
Record Types
A1
Schema
▪
Boxes represent record types
▪
Lines correspond to links
▪
Includes a data definition language (DDL) and a data
manipulation language (DML)
Rooted Trees
▪
The records are organized into forests (collections of rooted
trees).
▪
Dummy nodes are used for each tree root.
▪
A parent node can have multiple children (1:N).
▪
A child node has exactly one parent (1:1).
▪
No cycles are allowed in the structure.
A2
Dummy Node
for Records of
type B
B1
B2
B3
10
Hierarchical
Database Model
Hierarchical Database Model Continued
▪
Representing many to many (M:N)
relationships between two record types A
and B is accomplished through record
duplication.
▪
▪
▪
A one to many relationship from A to B (tree T1)
▪
A one to many relationship from B to A (tree T2)
Record duplication is necessary to preserve the
tree- structure organization of the database.
▪
Data inconsistency may result during updates
▪
Waste of space is unavoidable
Root of the tree T2
Root of the tree T1
A1
B1
Create two different trees to depict the one to
many relationships.
B1
A2
B2
B3
B2
A1
A1
B1
B3
A2
A2
11
Hierarchical
Database Model
Hierarchical Database Model Continued
▪
Addressing Data Duplication with Virtual
Records
▪
▪
Dummy Node for Records of type A
Contain no data, only represent a logical pointer
to a physical record.
When a record is to be replicated in several
database trees, only a single copy of the record is
kept in one of the trees. All other records are
replaced with virtual records.
A1
A2
Dummy Node
for Records of
type B
B1
Root of the tree T1
B3
Root of the tree T2
Virtual-B1
Virtual-A1
Virtual-B2
Virtual-B2
Virtual-B3
Virtual-A2
Virtual-A1
Virtual-B1
B2
Virtual-B3
Virtual-A2
Virtual-A1
Virtual-A2
Virtual-B1
12
Network
Database Model
Network Database Model
▪
A network database consists of a collection of
records which are connected to one another through
links.
▪
A record corresponds to a collection of fields, each field
contains a single data value.
▪
A record and its fields are represented by a record type.
▪
A link corresponds to an association between exactly two
records.
▪
Unlike in a hierarchical database, network databases allow
cycles and can accommodate arbitrary information graphs.
▪
▪
Schema
▪
Boxes represent record types
▪
Lines correspond to links
▪
Links can be one- to- one (1:1), one- to- many (1:N),
many- to- one (N:1), and many- to- many (M:N).
▪
Includes a data definition language (DDL) and a data
manipulation language (DML)
Examples
▪
Computer Associates Integrated Database Management
System (CA IDMS)
Graph which represents the relationship between A and B
Link
A1
Record Type A
Record Type B
B1
A2
B2
B3
13
Relational
Database Model
Relational Database Model
▪
A relational database consists of a
collection of tables (relationships).
▪
▪
Rows in each table represent individual
records.
Columns in each table represent attributes
(or fields).
▪
Each table is made up of key and non- key
fields.
▪ Associations between tables (relationships)
are realized through other tables
▪
Table that represents all records of type T
Attr1 Attr2 Attr3
Attrm-1 Attrm
...
Record1
Record2
.
.
.
Recordn
Examples
▪
▪
Apache Derby, IBM DB2, Informix, Ingres,
Microsoft Access, PostgreSQL
Microsoft SQL Server, MySQL, Oracle,
Paradox, JavaDB
Table for A
Table for B
Table for Relationship between A and B
14
Relational Database Model Continued
▪
▪
Relational Database Theory
▪
Based on the concept of normal forms.
▪
The higher the normal form for a table, the less
susceptible it is to inconsistencies and anomalies
ACID Properties
▪
Normal
Form
Relational
Database Model
Atomicity - All operations occur or none occur, no
partial transactions
▪
Consistency - Transaction brings the database
from one valid state to another valid state
▪
Isolation - No transaction should be able to
interfere with another transaction
▪
Durability - Once a transaction has been
committed, the changes are permanent
Normal
Form
3NF
All records have the same number of fields, no nested fields.
2NF
1NF and all fields in the key are needed to determine the values of
the non-key fields.
2NF and no non-key fields depend on any field(s) that are not the primary key.
EKNF
A subtle enhancement to 3NF for when there is more than one unique
composite key and keys do not have one or more fields in common.
BCNF
(Boyce-Codd Normal Form) 3NF and every determinant (field used to
determine another field in the table) could be a primary key.
4NF
A multi-valued dependency (MVD) is a functional dependency where the
dependency may be to a set and not just to a single value. It is defined as X→→Y
in relation R(X,Y,Z), if each X value is associated with a set of Y values in a way
that does not depend on the Z values.
BCNF and for every non-trivial multi-valued dependency (X→→Y) in F+ (closure of
functional dependencies), X is a super-key of R.
5NF
(PJNF)
(Project-Join Normal Form) A join dependency (JD) can be said to exist if the join
of R1 and R2 over C is equal to relation R; where R1 and R2 are the
decompositions R1(A,B,C) and R2(C,D) of a given relation R(A,B,C,D).
4NF and every join dependency is a consequence of its relation (candidate)
keys. That is, for every non-trivial join dependency *(R1,R2,R3) each
decomposed relation Rj is a super-key of the main relation R.
Description
1NF
Description
DKNF
6NF
(Domain-Key Normal Form) Requires that a table contain no constraints other
than domain constraints and key constraints.
Requires that the database table contain no non-trivial join dependencies. That is,
the table is in 5NF, is of degree n, and has no key of degree less than n - 1.
15
Relational Database Model Continued
▪
Relational
Database Model
Keys
▪
Simple
▪
▪
Single attribute that uniquely identifies
each tuple (row) in a table.
Primary
▪
▪
Unique set of attributes that identifies
each tuple (row) in a table.
Composite
▪
▪
Two or more attributes that uniquely
identify each row; where at least one
attribute is NOT a simple key on its own.
Compound
▪
Two or more attributes that uniquely
identify each row; where each attribute is
a simple key on its own.
▪
Candidate
▪
▪
A minimal super key.
Super Key
▪
▪
A set of attributes for a relation upon
which all attributes are functionally
dependent.
Foreign
▪
Unique set of attributes that identifies
each tuple (row) in a different table.
16
Relational
Database Model
Relational Database Model Continued
▪
▪
Data Manipulation
Structured Query Language (SQL)
▪
Select (Vertical/Horizontal Slicing), Update, Delete
▪
A declarative (as opposed to imperative),
standards based language (e.g. SQL- 2011)
for creating, querying, and manipulating
relational databases.
▪
Join (Building Intermediate Tables)
▪
Query Optimization
Data Definition
▪
Set Operations
▪
▪
▪
▪
Create, Alter, Drop
▪
Indexes, Constraints, Triggers, Stored Procedures
▪
Access controls
Cross, Theta, Equi, Natural, Inner, Full Outer, Left
Outer, Right Outer
In, Not In, Union, Intersect, Except (Difference),
Group By, Having,
▪
Nested Queries
▪
Views
Selection
Join
© 2013- 2014 Cisco and/ or its affiliates. All rights reserved.
Selection
➡
Cisco Confidential
22
Relational
Database Model
Relational Database Model Continued
R
R1
R2
R3
1
2
3
2
3
4
S
S1
S2
S3
3
4
5
1
2
3
Select *
From R,S;
Select *
From R cross join S;
T
R1
S1
1
3
3
1
Select *
From R,T
Where R.r3 < T.s1;
Examples
Select *
From R join T
On R.r3 < T.s1;
Select *
From R Left Outer Join T
On R.r1 = T.r1;
Select *
From R Full Outer Join T
On R.r1 = T.r1;
R1
R2
R3
S1
S2
S3
1
2
3
3
4
5
R1
R2
R3
R1
S1
R1
R2
R3
R1
S1
1
2
3
1
2
3
1
2
3
1
3
1
2
3
1
3
2
3
4
3
4
5
Equi Join (theta join using =)
2
3
4
Null
Null
2
3
4
1
2
3
Cross Join (cross product)
Select *
From R,T
Where R.r1 < T.r1;
R1
1
2
R2
2
3
R3
3
4
Select *
From R join T
On R.r1 < T.r1;
R1
3
3
S1
1
1
Left Outer Join (all rows from left)
Select *
From R natural join T;
R1
R2
R3
S1
1
2
3
3
R1
R2
R3
R1
S1
Select *
From S inner join T
on S.s3 > (T.r1 + T.s1);
1
2
3
1
3
Null
Null
Null
3
1
S2
S3
R1
S1
Union
(Select *
From R Right Outer Join T
On R.r1 = T.r1);
Select *
From R Right Outer Join T
On R.r1 = T.r1;
Natural Join (equi join on common attributes)
S1
(Select *
From R Left Outer Join T
On R.r1 = T.r1)
Right Outer Join (all rows from right)
R1
R2
R3
R1
S1
1
2
3
1
3
2
3
4
Null
Null
Null
Null
Null
1
3
Relational
Database Model
Relational Database Model Continued
R
R1
R2
R3
1
2
3
2
3
4
S
(Select r1 From R)
Union
(Select r1 from R1
T);
S1
S2
S3
3
4
5
1
2
3
T
R1
2
2
Union (unique rows from two tables)
S1
1
3
3
1
1
U1
U2
U3
1
1
1
1
1
2
1
2
3
1
2
4
2
1
1
2
1
2
2
1
3
Examples Continued
Select u1
From U
Group By u1
Having count(u2) > 2 AND
sum(u2) > 3 AND
sum(u3) > 5;
Difference (rows in first table but not in second)
U1
Select *
From R
Where r1 In (2,4,6);
1
Group By (grouping) and
(Select r1 From R)
Intersect
(Select r1 from T);
R1
U
(Select r1 From R)
Except
(Select r1 from T);
1
3
R1
R1
R2
R3
2
3
4
Set Inclusion
Select *
From S
Where s2 Not In (1,2,3);
Intersection (unique rows in both tables)
© 2013- 2014 Cisco and/ or its affiliates. All rights reserved.
S1
S2
S3
3 Set 4Exclusion
5
Having (operations on aggregates)
Select count(*)
From
(Select u1
From U
Group By u1
Having count(u2) > 2 AND sum(u3) > 4) as Temp;
Count
2
Nested Query
24
Relational
Database Model
Relational Database Model Continued
▪
View
▪
A saved query that represents a virtual table.
▪
Allows information hiding.
▪
The virtual table is populated at access time.
▪
Read- only access
▪
▪
Select ... From view_name …
Materialized View
▪
▪
▪
Create view view_name As
SQL_Query;
Create OR Replace View view_name As
SQL_Query;
Drop View view_name;
A saved query that represents a persistent (as
opposed to virtual) table.
Like a view with respect to
▪
Information hiding
▪
Read- only Access
Differences from a regular view
▪
Refreshed periodically (configurable).
▪
DDL syntax (e.g. create materialized view …)
▪
Not available with every RDBMS
Saved
Query
Definition
Virtual
Table
View
Actual
Table
Materialized View
20
Object Database Model
Object Database Model
▪
▪
ODBMS also known as Object- Oriented Database
Management Systems (OODBMS)
Examples
▪
▪
Object
Class
db4o, Caché, eXtremeDB, Perst, Objectivity/ DB,
ObjectStore, Versant Object Database, ObjectDB, VOSS
Person
Class
▪
Class (Template, like a cookie cutter)
▪
Properties (attributes) / Behaviors (actions/methods)
▪
Access/Visibility to properties and behaviors
Object (a cookie cut into the memory dough)
▪
▪
Encapsulation
▪
▪
An instance of a class
Storing an object’s properties and behaviors
together as part of the instance
Relationships
▪
Inheritance (Single, Multiple) / Inheritance Hierarchy
Properties
getID, setID
Behaviors
IS-A
Object- Oriented Concepts
▪
ObjectID
Employee
Class
SSN, Name, Birthdate
getSSN, setSSN, getName, setName,
getBirthdate, setBirthdate, getAge
IS-A
Org, Dept, Title, Mgr, EmployeeID,
HireDate
getOrg, setOrg, getDept, setDept,
getTitle, setTitle, getEmployeeID,
setEmployeeID, getMgr, setMgr,
getReportingHierarchy,
getDirectReports, getCoworkers,
getHireDate, setHireDate
Properties
Behaviors
Properties
Behaviors
OODBMS are integrated with an object-oriented
programming language similar to RDBMS but with
an object-oriented database model. Objects, classes,
and inheritance are directly supported in the database
schemas and in the query language.
21
Object Database Model
Object Database Model Continued
▪
Object- Oriented Programming Languages
▪
▪
▪
C++, Java, C#, JavaScript, Ruby, Smalltalk, Scala, Groovy,
ParaSail, Ceylon, Clojure, JRuby, ...
Object- Oriented Applications
▪
Dynamically create and destroy objects
▪
Leverage an Object Graph during the application’s execution
▪
Transactions
▪
Queries
▪
Indexes
▪
Administration, including tuning
Object- Oriented Database Management Systems
▪
Support the modeling and creation of data as objects
▪
Include support for classes of objects and the inheritance of
class properties and behaviors (methods) by subclasses and
their objects.
▪
Create, Read (Search), Update, and Delete objects in the
Database <CRUD Operations>
▪
The class structure is the database schema
▪
Persistence - Explicit and Transparent
▪
Explicit Persistence - CRUD operations are performed in the code
▪
Transparent Persistence - Objects are moved to and from the
database invisibly
Instantiated Objects @ Time t
22
Object- Relational Database Model
▪
Try to bridge the gap between
traditional RDBMS and OODBMS
▪
Includes the full suite of RDBMS features
▪
Object- oriented features typically vary by
vendor and revolve around the SQL- 99
specification
Inheritance (Table & Type)
User- defined Data Types (Attributes &
Tables)
Functions for user- defined data types (UDTs)
▪
▪
▪
▪
Object Relational
Database Model
Examples
▪
PostgreSQL, CUBRID, Oracle, Informix, DB2,
SQL Server
23
Object- Relational Database Model Continued
▪
A popular alternative to an ORDBMS is
using an Object Relational Mapping (ORM)
framework with a RDBMS
▪
▪
Object Relational
Database Model
Objects
Object-Oriented
Application
Apache Cayenne, Hibernate, JDO, JPA,
GORM, Active Record, ...
ORM frameworks allow
▪
Software engineers to focus on and work
with objects
▪
Database designers to focus on and work
with relational database constructs
▪
The impedance mismatch between
objects and tables to be transparently
handled
Maps objects to
tables and vice
versa
ORM Framework
RDBMS
Tables
24
XML Database Model
▪
Two approaches: XML- enabled and Native
XML databases
▪
XML- enabled databases
▪
▪
rely on a middle- tier to transform XML to another
DB representation
Native XML Databases
▪
▪
XML Database Model
store, manipulate, and query XML documents
Examples
▪
Sedna, Xindice, BaseX, eXist, MarkLogic Server,
MonetDB/ XQuery
25
Content Management
Systems
Content Management Systems
▪
Enterprise Content Management Systems (ECMs)
▪
Provide a mechanism to organize documents in various
formats (structured and unstructured data)
▪
Administrative and User Tools
▪
Access control based on roles and permissions
▪
Storage and retrieval of data/ Version Control
▪
Workflows
▪
▪
APIs
▪
Proprietary
▪
JSR- 170 (Content Repository API for Java)
▪
Content Management Interoperability Services
(CMIS)
▪
Open standard for controlling diverse document
management systems & repositories
Examples
▪
OpenCMS,
Alfresco,
WordPress,
Apache Lenya,
Apache Jackrabbit,
SharePoint,
Interwoven,
Documentum,
User
Application
User + Admin Tools
API
Content
Content Management System
26
Data Warehouse
Enterprise Data Warehouse
▪
A giant data repository facilitating
various types of data aggregation,
reporting, business intelligence (BI),
data mining, and analytics processing.
▪
▪
Data marts represent specialized
data warehouses.
Tools are leveraged to extract new
insights from the data warehouse
and data marts.
Sales
Marketing
ETL
ETL
Supply
Chain
ETL
Operations
ETL
Data Vault
…
▪
Data from the various source
systems is placed into the
warehouse via extract, translate, and
load processes (ETL).
Data
Warehouse
Data Sources
ETL
Exploration,
Mining, and
Reporting
Tools
Data
Mart(s)
27
Distributed Databases
Distributed Databases
▪
Database system in which the data - and
sometimes processing – is not centralized
▪
Database Duplication
▪
Active/Passive Configuration
▪
Database Replication
▪
Active/Active Configuration
▪
Conflict detection and resolution
▪
Database Fragmentation (DRDBMS)
▪
Data distributed/partitioned across locations
▪
Vertical (columns) and horizontal (rows) slicing
▪
Semi- Joins
▪
Expensive to ship data across the network
▪
Local query optimization based on costs/combine results
28
Part III:
New methods for Storing Big Data
29
Not Only SQL/No SQL (NOSQL) Approaches
▪
NOSQL databases represent newer data models
aimed at Big Data problems.
▪
Differ from the relational model in several ways:
▪
▪
SQL is not used as the primary query language
▪
Fixed- table schemas may not be required
▪
Joins are generally not supported
▪
ACID (atomicity, consistency, isolation, durability) guarantees
may not exist
▪
Architectures typically leverage massively distributed
computing resources (processing and storage)
CAP Theorem (Brewer’s Theorem)
▪
Impossible for a distributed computer system to
simultaneously provide three of the following guarantees:
▪
Consistency - All nodes see the same data at the same time
▪
Availability - Every request receives a response about whether
or not it was successful
▪
Partition Tolerance - The system continues to operate despite
arbitrary message loss or failure of part of the system
30
Not Only SQL (NOSQL) Approaches Continued
▪
NOSQL Database Types
▪
▪
Key- Value Stores
▪
Document Stores
▪
Column Stores
▪
Graph Databases
Sharding
▪
Horizontal Partitioning
▪
Breaking a large database into several smaller
databases that share nothing
▪
▪
The smaller databases can be distributed across
multiple servers.
Size of database and # of transactions increases
linearly, while query response time increases
exponentially
Shard 2
Large
Database
Partitioning scheme
(e.g. hash function)
…
▪
Shard 1
Categorized according to the way they store their
data
Application
Shard N- 1
Shard N
31
Key- Value Stores
▪
▪
Keys are unique
▪
Values do not have to be unique
Basic operations:
Key
Value
▪
AddOrUpdate(key, value)
keyi
valuei
▪
GetValue(key)
keyj
valuej
▪
DeleteKey(key)
▪
DoesKeyExist(key)
Feature Differentiation
▪
Complexity of the keys
▪
Advanced operations (e.g. Expire, Lists, Sets, Hashes, …)
▪
Distributed vs. Non- distributed
▪
Memory- resident vs. Disk- based
…
▪
▪
…
▪
Two- column table (key, value)
keyn-1
valuen-1
keyn
valuen
Examples
▪
Redis, Voldemort, Riak, Hibari, MemcacheDB, BerkeleyDB,
Amazon S3, …
32
JSON Syntax
Document Stores
▪
▪
▪
A database of JavaScript Object Notation
(JSON) or Binary JSON (BSON) “ documents”
▪
JSON is a light- weight data interchange format
▪
Based on a subset of the Object- Oriented
JavaScript programming language
Documents are analogous to records with
fields and values.
▪
Grouped into collections.
▪
Collections are indexed.
“name” : “Michael”,
“GolfHandicapIndex” : 5,
“Scores” : [
{“course” : “Lakeridge”,
“score” : 73},
Examples
▪
▪
{
CouchDB, MongoDB, RavenDB, …
Programming Language Tools &
Frameworks
{“course” : “Wolf Run”,
“score” : 77},
▪
{”course” : “Wildcreek”,
“score” : 79}
See http://json.org
From json.org
]
}
JSON Example
33
Bigtable
ColA ColB ColC … ColM
Column Stores
▪
Motivated by Google’s “ Bigtable: A Distributed
Storage System for Structured Data” [2006]
▪
For random read/ write access to big data – consisting of
billions of rows and millions of columns – atop clusters of
commodity hardware.
▪
Vertical partitioning of the data according to the attributes
(columns).
▪
“ A Bigtable is a sparse, distributed, persistent, multidimensional sorted map [row key, col key, timestamp]
M,N are
large
Row N
Main Ideas
▪
Large tables can be expensive to process (entire row
must be read).
▪
Extensible records that are partitioned across nodes.
▪
Rows and columns comprise the data model.
▪
Horizontal sharding based on row keys (key ranges)
▪
Columns can be partitioned into column groups/column
families (allow related columns to be kept together)
timestamp
…
▪
Row 1
Row 2
Bigtable
Apache
HBase
Apache
Accumulo
Apache
Cassandra
DynamoDB,
Hyberbase, …
34
Column Stores Continued
{
Assume the columns are
aggregated into 5 different
column groups/families
ColA ColB ColC … ColM
{
M,N are large
…
Assume the row keys
are partitioned into 4
different ranges
Row 1
Row 2
timestamp
Row N
CF1
CF2
CF3
CF4
CF5
timestamp
Row key range 1
Row key range 2
Row key range 3
Row key range 4
Database distributed over 4 x 5 = 20 nodes
35
Graph Databases
▪
▪
Basics
▪
Leverage graph structures with nodes, edges, and
properties to represent and store data.
▪
Nodes in a graph are similar to objects in that they have
attributes/properties
▪
Edges are used to represent a relationship between two
nodes or between a node and a property
▪
Properties represent attributes that are associated with
nodes or edges (relationships)
▪
Hyper-e dges represent a relationship between a set of
nodes.
▪
A traversal navigates a graph, equivalent to performing a
query
▪
Fully- t ransactional, enterprise-s trength databases
Application developers leverage an objectoriented, flexible network structure instead of
static tables
Undirected Edge
Node
{
{
A
C
B
Property1 : Value 1
Property2 : Value 2
…
PropertyN: Value N
PropertyA : Value 1
PropertyB: Value 2
…
Property Z: Value Z
Undirected Graph
E
Node
D
Directed Edge
F
G
Directed Graph
41
Graph Databases Continued
▪
Key Points
▪
▪
▪
▪
▪
▪
Property Graphs
Query = Traversal
Network Science
Graphs are everywhere!
Examples
▪
Neo4j, InfiniteGraph, OrientDB, AllegroGraph,
Titan, …
Person
Name: X
ID: Y
…
P
interested- in
has-role
Nodes (vertices) represent entities.
▪ Edges represent relationships.
▪ Nodes and edges are able to have
properties
follows
YearsInRole: Y
Ratings: R
…
Areas of Interest
Interest Area: I
…
IA
EA
J
requires
RequiredProficiencyRating: W
…
Job Roles
Job Role: J
…
Areas of Expertise
Expertise Area: E
…
Queries/ Traversals/ Algorithms for identifying:
Who are the best mentors for a person P? Who
are the experts in expertise area E? What
areas does person P need to improve in? What
are the intellectual capital risk areas?
How do the people in job role J compare / who is promotion ready?
.
.
.
42
New SQL Approaches
▪
Motivation
▪
▪
High- volume transaction- oriented systems (e.g. financial,
order processing, etc.) cannot give up strong
transactional and consistency requirements, and
therefore are left wanting with respect to NOSQL options.
Hybrid Architectures
▪
A class of RDBMS that aim to provide the same scalability
as NOSQL systems for on- line transaction processing
(OLTP) –significant read/write activity – whilst still
maintaining ACID properties and utilizing SQL as the
primary interface.
▪
▪
Parallel Databases (parallelization of
various operations)
▪
Multi- processor Architectures
▪
Shared Memory – multiple processors share the main
memory space
▪
Shared Disk – nodes have autonomous memory, but share
mass storage
▪
Shared Nothing – nodes have their own main memory and
mass storage
▪
Memory access relative to the processor (local vs. other vs.
shared)
Cluster
▪
Approaches
▪
Non- Uniform Memory Access (NUMA)
▪
Enter New SQL
▪
▪
▪
Distributed cluster of shared- nothing nodes where each node
owns a subset of the data. Transactions and queries are
fragmented and routed to the nodes that contain the needed
data.
In- Memory Databases
▪
Memory is always faster than mass- storage
▪
For high- volume transactions that are short- lived,
access a small subset of the available data, and are
executed over and over with different inputs.
NewSQL Examples
▪
Google Spanner, Clustrix, FoundationDB, NuoDB,
Translattice, VoltDB, Pivotal’s GemFire & SQLFire,
MemSQL
43
Data Virtualization
▪
▪
▪
www.cisco.com/ web/ services/ enterprise- it- services/ data- virtualization/ documents/ cisco- information- server- 62- ds.pdf
An approach to data management that allows an
application to retrieve and manipulate data without
requiring details about the underlying data model or
where the data is located.
▪
Differs from ETL in that the source data remains in place.
▪
Data in the source systems is readable and can also be
writable.
Features
▪
Abstraction (location, data model, API, access language, …)
▪
Virtualized Data Access (common access point)
▪
Transformation (transform/reformat for use)
▪
Data Federation (combine data sets from multiple sources)
▪
Data Delivery (publish views and/ or services for reuse)
Examples
▪
Denodo, Composite (Cisco), Informatica, IBM SmartCloud
Data Virtualization, …
44
http:/ / www.denodo.com/ en/ product/ features.php
Cisco Confid
Part IV:
Tools for Processing and Accessing
Big Data
The Big Data Tool Zoo
46
The Big Data Tool Zoo (Part 1): Apache Hadoop
▪
▪
Framework which enables distributed processing
of large data sets across clusters of computers.
Primary Components
▪
Hadoop Common
▪
Hadoop Distributed File System (HDFS)
▪
Build an HDFS instance
▪
Use FS commands for interactive access
▪
▪
▪
▪
hadoop fs –copyFromLocal myBigData / user/ hadoop/ demo
▪
hadoop fs –cat / user/ hadoop/ demo/ myBigData
▪
Many other commands
Facilitates interaction patterns for data in HDFS
▪
Batch (MapReduce), Interactive (Tez), Online (HBase)
▪
Streaming (Storm), Graph (Giraph), In- memory (Spark), others
Hadoop MapReduce
▪
©\
hadoop fs –mkdir –p / user/ hadoop/ demo
Hadoop YARN
▪
Hadoop MapReduce
+ Batch- oriented, parallel processing of
large data sets
Batch- oriented processing: Map, Shuffle, Reduce
+ Processing large
data sets.
Hadoop YARN
+ Framework for job scheduling and cluster resource management.
+ Facilitates broad array of data interaction patterns.
Hadoop Distributed File System (HDFS)
hadoop fs –ls
▪
Other
Tools
+
+
+
+
+
+
+
+
+
+
Redundant, reliable storage.
Designed to run on commodity hardware.
Highly fault tolerant (failures expected)
Fast fault detection and automatic recovery.
Suitable for large data sets distributed across multiple nodes.
Provides high aggregate data bandwidth.
Designed for batch processing and providing streaming access.
Scales to hundreds of nodes in a single cluster.
Supports tens of millions of files in a single instance.
Interactive access provided via File System (FS) commands.
Hadoop Common
+ Common utilities and libraries to support the other modules.
Apache Hadoop Ecosystem
47
MapReduce
▪
▪
MapReduce
▪
Google paper [2004]
▪
High- level programming model and implementation for
large- scale parallel data processing
▪
Free variant: Hadoop
▪
Google claims in 2014 that its Cloud Dataflow is meant
to replace MapReduce.
MapReduce Programming Model
▪
Input and Output: each a set of (key, value) pairs
▪
Programmer specifies two functions:
▪
▪
Input Data
+ Processed in parallel in order to elicit a set of (key, value) pairs
Map Phase
+ Map(inKey, inValue) € List (outKey, intermediateValue)
+ System applies the map function in parallel to all input key/ value
pairs in the input file.
Shuffle Phase
+ All pairs with the same intermediate key are grouped together –
similar to what an SQL “ group by” would do
Map (inKey, inValue) € List (outKey, intermediateValue)
▪
Processes input key/value pair
▪
Produces (emits) a list of intermediate key/value pairs
Reduce (outKey, List (intermediateValue)) € List (outValue)
▪
Combines all intermediate values for a particular key
▪
Produces a set of merged output values (usually just one)
Reduce Phase
+ System applies the reduce function in parallel to all intermediate
values for a particular key and produces a set of merged output
values as a result
MapReduce FundamentalsCisco Confidential
48
MapReduce Examples
Word Frequency from a Corpus of Documents
(doc_id, value)
(word, count)
(word, list of values)
(word, frequency)
(id1,v1)
(w1,1)
(w1,(1,1,…))
(w1, 15)
(id2,v2)
(w2,1)
(w2,(1,1,…))
(w2, 27)
(w3,1)
(w3,(1,…))
(w3, 22)
(w1,1)
…
…
(id3,v3)
(w2,1)
…
How many times does
each unique word occur?
Reduce
…
Shuffle
Map
Word Length Histograms from a Corpus of Documents
(doc_id, value)
(size, count)
(size, list of values)
(size, frequency)
(id1,v1)
(small,7)
(small,(7,10))
(small, 17)
(id2,v2)
(medium,15)
(medium,(15,8))
(medium, 23)
(big,3)
(big,(3,4))
(big, 7)
Reduce
(small,10)
(medium,8)
How many big, medium,
and small words occur,
where big € 12+ letters,
medium € 5…9 letters, and
small € 1..4 letters?
(big,4)
Map
Shuffle
49
MapReduce Try for Yourself with jsmapreduce
1. Go to www.jsmapreduce.com
2. Register for a free account
3. Try the examples (JavaScript
and/ or Python)
4. Extend the examples or
experiment with your own data
Execution Controls
Input data
Map Function
Status
Reduce Function
Output
44
Cisco
JSMapreduce Example – Add 4- card straight to Poker
45
The Big Data Tool Zoo (Part 2): A broader perspective
Distributed application development framework; facilitates generic cluster resource management.
Application execution framework for complex directed- acyclic- graph (DAG) data
processing tasks; accelerates Hadoop query processing.
Bulk Synchronous Parallel(BSP) Computing; advanced analytics
beyond MapReduce; network algorithms, graph algorithms, machine
learning, ….
Data warehouse software allowing the querying and managing of large datasets
which reside in distributed storage using an SQL- esque language called HiveQL.
Also allows custom mappers & reducers. Includes HCatalog table storage/mgmt.
Workflow scheduler (MR, Pig, Hive, …)
Analysis of large data sets; parallelization of MR tasks; Pig Latin language
Iterative graph processing which extends Google’s Pregel.
Provision, manage, and monitor Hadoop clusters
Real- time distributed processing of incoming data streams;
real- time analytics, machine learning, …
Bulk data xfer between Hadoop and structured data stores.
Bigtable- esque (column store)
SQL- supported big data warehouse system for Hadoop
In- memory analytics, 100X faster than MapReduce;
general purpose data processing for large datasets.
Combines SQL (SparkSQL), streaming, and
complex analytics.
Fault- tolerant
Bigtable- esque
distributed database
MapReduce
Hive
Pig
Cassandra
Mahout
Monitoring
Chukwa
Drill
Samza
Tajo
Sqoop
Ambari
Twill
Hama
Oozie
MapReduce
(batch exec. f/w)
Giraph
Hive
Storm
Pig
Machine Learning
Streaming Event Data
HBase
ZooKeeper
Distributed query engine;
extends Google’s Dremel.
Flume
Spark
Distributed
Configuration
Mgmt.
Distributed stream processing, leverages Kafka messaging f/w
Tez
Hadoop YARN (Job Scheduling & Cluster Resource Management)
Hadoop Distributed File System (HDFS): Redundant, Reliable, Storage
Hadoop Common (Utilities and Libraries)
46